Pith. sign in

REVIEW 6 minor 50 references

This paper proves that a parameterized matrix family is fully determined up to a constant orthogonal gauge by its pointwise spectrum together with loop products of derivative couplings, and that this inversion is Lipschitz stable.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-03 15:10 UTC pith:D4SZPHBF

load-bearing objection A genuinely new complete, reduced, and stable pointwise invariant for orthogonal gauge classes; the mathematics is sound and the soft spots are the connectedness assumption and thin reproducibility.

arxiv 2607.29021 v1 pith:D4SZPHBF submitted 2026-07-31 math.NA cs.NA

Stable Recovery of Matrix Gauge Classes from Pointwise Invariants

Dexuan Zhou , Huajie Chen , Bernie Hsu , Christoph Ortner This is my paper
classification math.NA cs.NA MSC 15A2915A1815A2105C2205C50
keywords gauge equivalenceinverse spectral problemsderivative couplingBerry connectionloop productsHamiltonian learningmatrix familiessign gauge
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper tackles the inverse problem: which basis-independent observations of a configuration-dependent matrix family H(x) identify it up to a global change of basis? It proves that the pointwise spectrum alone is insufficient, but augmenting it with products of derivative-coupling entries along cycles in the instantaneous eigenframe yields a complete invariant. Under non-degeneracy and connectivity of the interaction graph, two families with identical invariants are necessarily gauge equivalent (Theorem 2.1). On compact domains, the inversion is Lipschitz stable: small invariant discrepancy implies small equivalence distance (Theorem 3.1). The reduced invariant records O(M N^2) scalars per configuration and is numerically practical.

Core claim

The central discovery is that the gauge class of a non-degenerate C^1 family H: X -> Sym(N) is characterised exactly by pointwise data: the ordered spectrum, squares of the derivative-coupling entries (pure 2-loops), mixed 2-loops aligning signs across parameter directions, and products along a spanning-tree-reduced set of fundamental cycles. Equality of these data for two families forces them to be related by a single constant orthogonal matrix, provided the aggregate interaction graph is connected at every configuration. The proof reconstructs the coupling matrices up to vertex sign gauges, shows the signs have trivial cycle holonomy, and then glues locally constant orthogonal gauges into

What carries the argument

The central object is the derivative coupling (Berry connection) A^mu(x;H) = u(x)^T d_{x_mu} u(x), with u an instantaneous eigenframe. Loop products along cycles of the interaction graph—products of coupling entries around closed loops—are invariant under the residual sign gauge that flips eigenvector signs. A spanning-tree reduction selects O(|E|) fundamental cycles, and an active-direction selection reduces the mixed 2-loops to O(M|E|). Together with the spectrum, these assemble into the invariant P_{T,ell}(x;H) that the completeness and stability theorems hang on.

Load-bearing premise

The coupling graph between instantaneous eigenstates must remain connected at every configuration; if it disconnects at a point that separates the domain, incompatible local gauges can be glued across the gap and the invariant no longer determines the global gauge class.

What would settle it

Build two C∞ families on a connected interval that share every pointwise invariant—same spectrum, pure and mixed two-loops, and fundamental-cycle products—but are not related by any constant orthogonal matrix, for instance by flipping the sign of the off-diagonal coupling on one side of a point where the coupling vanishes and the graph disconnects (as in the paper's Examples 3.4–3.5). Agreement of all invariant components for such a pair would refute the completeness theorem.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Hamiltonian learning can be performed from basis-independent observables, decoupling the basis of the training data from the basis of the surrogate model.
  • The invariant costs O(M N^2) scalars per configuration and evaluates linearly in the number of sampled configurations, making it practical for fitting.
  • On compact domains with uniform spectral gap, bounded couplings, and a lower-bounded active tree connection, the gauge class is recovered stably: dist_K <= C Delta_K.
  • Every component of the invariant is necessary: dropping the spectrum, pure 2-loops, mixed 2-loops, or fundamental cycles destroys completeness in the relevant regimes.
  • Exceptional sets—spectral crossings or graph disconnections—are removable when they are non-separating or the family is analytic, so generic systems remain identifiable.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same loop-product mechanism should extend to complex Hermitian families under unitary gauge, where the residual eigenframe freedom is a continuous U(1)^N group and second-order (curvature) information would enter the invariant; the paper notes this as separate work.
  • The pointwise loop-product invariant is equivalent to the classical continuous trace words on the admissible class, but much cheaper to evaluate; this suggests trace-word-type invariants can be replaced by pointwise data in practice.
  • In the large-system limit where N and M both grow with particle number, sparsity of the interaction graph could reduce the O(M N^2) cost; the paper leaves this as the key open problem.
  • The invariant doubles as a quotient distance for comparing reduced-basis surrogate models without fixing a basis, which could complement model-order-reduction pipelines.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

0 major / 6 minor

Summary. The paper studies the inverse problem of identifying a C^1 family of real symmetric matrices H(x), up to constant orthogonal conjugation (gauge equivalence), from gauge-invariant observations. It constructs an invariant P_{T,ℓ} consisting of the spectrum, squared derivative-coupling entries (pure 2-loops), mixed 2-loop products on interaction-graph edges, and fundamental-cycle products with respect to a spanning tree of the pointwise interaction graph. Theorem 2.1 proves completeness of this invariant under non-degeneracy and pointwise connectedness of the aggregate interaction graph. Theorem 3.1 proves a Lipschitz stability estimate on compact, rectifiably path-connected subsets, with an explicitly stated smallness threshold. The paper also establishes necessity of each invariant component and each structural hypothesis, gives removable-exceptional-set results, and reports numerical optimization experiments.

Significance. If the results hold, they provide a computationally practical complete invariant for a gauge-equivalence problem arising in Hamiltonian learning. Unlike multi-point trace words, the invariant is pointwise, records O(MN^2) scalars per configuration, and is linearly evaluable in the number of sampled configurations. The proof of Theorem 2.1 is detailed and internally consistent; the local sign reconstruction, the connected-graph gluing step, and the avoidance of topological sign phases are all handled carefully. The stability proof in Appendix B is elaborate but sound: the smallness threshold is used precisely to transfer spectral gaps, preserve tree activation, and select the correct discrete sign branch. The counterexamples showing necessity of each invariant component and of the connectedness assumption are valuable, as is the analytic-removability discussion. The main limitation is the pointwise connectedness assumption, but the paper acknowledges it explicitly and gives both necessity examples and extensions via exceptional sets and analyticity. The absence of code/data and some presentation inconsistencies affect reproducibility but not the correctness of the central mathematic

minor comments (6)
  1. [Abstract; §2.1] The abstract states that the pointwise spectrum is incomplete 'already for linear families on R', but the illustrative example (2.1) uses H2(x)=R(x)^T H1 R(x) with a rotation by angle x, which is not a linear-in-x family. Please either provide a genuine linear counterexample or soften the claim.
  2. [Table 5] The text in §3.2 says that the mixed 2-loops 'adds nothing here' for Example 3.6, but Table 5 shows +2-loops distances considerably worse than +2p-loops for (M,N)=(2,4) (1.6e-5 vs 6.3e-7) and (3,4) (9.4e-4 vs 1.1e-6). Please clarify whether this is optimizer variance or an actual effect, and correct the text if needed.
  3. [Remark A.1] The assertion that the trace tr(P_{m1} ∂_{x_{µ1}}H ··· P_{mk} ∂_{x_{µk}}H) is 'a fixed parameter-derivative' of the multi-point trace word T_k evaluated at x_1=···=x_k=x is not immediate. Differentiating T_k at the diagonal produces sums over all factors and does not obviously isolate the spectral projectors P_{m_r}. Please supply the explicit differentiation formula or revise the remark.
  4. [§3 (numerical experiments)] No code or data are shipped, and the tables report only aggregate values. Since the experiments use 100 restarts and discard high-loss restarts, please report restart counts, seeds, and the distribution of distances/RMSEs (or at least error bars) to make the claims reproducible.
  5. [Table 4] The table header labels the invariant set as '+2p-loops', but the text says the optimization is run on the complete invariant. For N=2, M=1 the pure 2-loops indeed complete the invariant, but the label is potentially misleading; please state this equivalence explicitly.
  6. [Figure 2] The axis labels and subplot structure in Figure 2 are difficult to read in the text; please redraw with clearer labels and, if possible, use a consistent scale for RMSE and dist.

Circularity Check

0 steps flagged

No significant circularity; the completeness and stability proofs are self-contained and the invariant is not a fitted or self-referential quantity.

full rationale

I walked the derivation chain from Definition 2.1 through Theorem 2.1, the sharpness examples, and the stability proof in Section B. The invariant P_{T,ℓ} is constructed directly from the spectrum, the derivative couplings, and loop products of H itself; it is not defined in terms of the target gauge class, and no parameter is fitted to the data that the theorems then 'predict'. Theorem 2.1's converse proof is self-contained: Steps 1–6 reconstruct the coupling matrices up to a sign gauge from equality of the invariant components, use cycle-holonomy and signed-graph balancing to factor edge signs into vertex signs, and glue local gauges using the connected interaction graph. The gluing argument never assumes H1 ∼ H2; instead it derives the constant orthogonal gauge from invariant equality. Theorem 3.1 is a genuine stability theorem proved through Lemmas B.1–B.6, with constants depending only on the spectral gap, coupling bound, tree-activation modulus, and path-length bound; the inequality dist_K ≤ C Δ_K is a proven modulus relating two a priori different quantities, not a restatement of the definition of Δ_K. The relationship to continuous trace words is handled honestly in Appendix A: the authors do not claim to invent completeness from scratch, but prove their pointwise loop-product invariant is complete and stable, and show explicitly how the two invariant families determine one another. Self-citations in the introduction (e.g., Barrett et al. 2025) are motivational and not load-bearing; the graph-gauge principle cited from Korotyaev and Saburova is reproved in Proposition 2.2 and used with a full proof of the cycle-basis argument. The limitations flagged in Section 4 and Remark 3.2—theoretical scope, no released code, real-symmetric restriction, and the genericity of the codimension-two exceptional set—are honest scope restrictions, not hidden circularity. No step reduces by construction to its own input, and no central claim depends on an unverified self-citation.

Axiom & Free-Parameter Ledger

0 free parameters · 7 axioms · 0 invented entities

No free parameters are fitted in the main theorems; the numerical experiments use trainable model parameters that are not part of the analytical claim. The axioms are standard spectral/graph-theoretic facts plus explicitly stated domain assumptions. No new physical entities are introduced.

axioms (7)
  • domain assumption X ⊂ R^M is open and connected.
    Connectedness is required for global gluing of local gauges in Theorem 2.1; disconnected domains break it (Examples 3.4–3.5).
  • domain assumption H is C^1 and non-degenerate (simple spectrum) on X.
    Simple spectrum is needed for C^1 ordered eigenframes and the derivative-coupling formula (Section 2.2).
  • domain assumption The aggregate interaction graph G(A(x;H)) is connected at every x ∈ X.
    Connectivity is used in Step 6 of Theorem 2.1 to force the residual gauge to be ±I and in the stability modulus (3.4).
  • standard math Kato's eigenframe regularity: simple spectrum yields C^1 ordered eigenframes on simply connected neighbourhoods.
    Invoked when defining derivative couplings and local gauges in Sections 2.2 and Appendix B.
  • standard math Fundamental cycles span the cycle space over F2; balanced signed graphs have vertex signs.
    Used in Step 3 of Theorem 2.1, citing Bollobás and Harary.
  • standard math Identity theorem for real-analytic functions.
    Used in Remark 3.1 to extend local gauges across exceptional sets for analytic families.
  • domain assumption For Theorem 3.1: uniform spectral gap g_K > 0, finite B_K, and active-tree lower bound κ_K > 0 on compact K.
    Quantitative stability requires these moduli; they are constructed in Remark 3.4 and used throughout Appendix B.

pith-pipeline@v1.3.0-daily-deepseek · 35484 in / 14904 out tokens · 155004 ms · 2026-08-03T15:10:40.625843+00:00 · methodology

0 comments
read the original abstract

A parameterized matrix family $x\mapsto H(x)$ on a configuration domain is determined by its physical content only up to a constant orthogonal change of basis. This gauge ambiguity is intrinsic to data-driven Hamiltonian models, such as tight-binding parameterizations, reduced-order electronic structure methods, or excited-state models. It raises a basic inverse problem: what observations of $H(x)$ suffice to identify the family up to this gauge? The pointwise spectrum is incomplete already for linear families on $\mathbb{R}$. Here, we prove that, under natural non-degeneracy and connectivity assumptions, augmenting the spectrum with loop products of the coupling matrices in the instantaneous eigenframe yields a complete invariant and that inversion is stable. We support the theory with numerical experiments.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

50 extracted references

  1. [1]

    Maximally localized

    Marzari, Nicola and Mostofi, Arash A and Yates, Jonathan R and Souza, Ivo and Vanderbilt, David , journal=. Maximally localized

  2. [2]

    2025 , eprint =

    Manucci, Mattia and Stamm, Benjamin and Zeng, Zhuoyao , title =. 2025 , eprint =

    2025

  3. [3]

    2026 , eprint =

    Stamm, Benjamin and Zeng, Zhuoyao , title =. 2026 , eprint =

    2026

  4. [4]

    2024 , eprint =

    Garrigue, Louis and Stamm, Benjamin , title =. 2024 , eprint =

    2024

  5. [5]

    Michigan Mathematical Journal , volume=

    On the notion of balance of a signed graph , author=. Michigan Mathematical Journal , volume=

  6. [6]

    2024 , eprint =

    Manucci, Mattia and Mengi, Emre and Guglielmi, Nicola , title =. 2024 , eprint =

    2024

  7. [7]

    Russian Mathematical Surveys , volume=

    On matrices depending on parameters , author=. Russian Mathematical Surveys , volume=

  8. [8]

    Machine learning method for tight-binding

    Wang, Zifeng and Ye, Shizhuo and Wang, Hao and He, Jin and Huang, Qijun and Chang, Sheng , journal=. Machine learning method for tight-binding

  9. [9]

    Deep-learning density functional theory

    Li, He and Wang, Zun and Zou, Nianlong and Ye, Meng and Xu, Runzhang and Gong, Xiaoxun and Duan, Wenhui and Xu, Yong , journal=. Deep-learning density functional theory

  10. [10]

    General framework for

    Gong, Xiaoxun and Li, He and Zou, Nianlong and Xu, Runzhang and Duan, Wenhui and Xu, Yong , journal=. General framework for

  11. [11]

    and Maurer, Reinhard J

    Qian, Chen and Vitartas, Valdas and Kermode, James R. and Maurer, Reinhard J. , title =. npj Computational Materials , volume =. 2026 , doi =

    2026

  12. [12]

    Equivariant analytical mapping of first principles

    Zhang, Liwei and Onat, Berk and Dusson, Genevi. Equivariant analytical mapping of first principles. npj Computational Materials , volume=

  13. [13]

    npj Computational Materials , volume=

    Deep-learning atomistic semi-empirical pseudopotential model for nanomaterials , author=. npj Computational Materials , volume=

  14. [14]

    Electrical impedance tomography and

    Uhlmann, Gunther , journal=. Electrical impedance tomography and

  15. [15]

    Journal of the Society for Industrial and Applied Mathematics, Series A: Control , volume=

    Mathematical description of linear dynamical systems , author=. Journal of the Society for Industrial and Applied Mathematics, Series A: Control , volume=

  16. [16]

    Uniqueness in an inverse boundary problem for a magnetic

    Krupchyk, Katsiaryna and Uhlmann, Gunther , journal=. Uniqueness in an inverse boundary problem for a magnetic

  17. [17]

    Universal materials model of deep-learning density functional theory

    Wang, Yuxiang and Li, Yang and Tang, Zechen and Li, He and Yuan, Zilong and Tao, Honggeng and Zou, Nianlong and Bao, Ting and Liang, Xinghao and Chen, Zezhou and Xu, Shanghua and Bian, Ce and Xu, Zhiming and Wang, Chong and Chen, Si and Duan, Wenhui and Xu, Yong , journal=. Universal materials model of deep-learning density functional theory

  18. [18]

    Vanderbilt, David , year=

  19. [19]

    Nature , volume=

    Topological quantum chemistry , author=. Nature , volume=

  20. [20]

    Freiling, Gerhard and Yurko, Vjacheslav Anatoljevich , year=. Inverse

  21. [21]

    Inverse problems for matrix

    Yurko, V , journal=. Inverse problems for matrix

  22. [22]

    2012 , publisher=

    Graph Theory: An Introductory Course , author=. 2012 , publisher=

    2012

  23. [23]

    Kato, Tosio , title =

  24. [24]

    Physical Review Research , volume=

    Calculating nonadiabatic couplings and Berry's phase by variational quantum eigensolvers , author=. Physical Review Research , volume=. 2021 , publisher=

    2021

  25. [25]

    A Characterisation of Smooth Maps into a Homogeneous Space , author=. SIGMA. Symmetry, Integrability and Geometry: Methods and Applications , volume=. 2022 , publisher=

    2022

  26. [26]

    Hans Hellmann: Einf

    Hellmann, Hans , year=. Hans Hellmann: Einf

  27. [27]

    Mathematics of the USSR-Izvestiya , volume=

    Trace identities of full matrix algebras over a field of characteristic zero , author=. Mathematics of the USSR-Izvestiya , volume=

  28. [28]

    Siberian Mathematical Journal , volume=

    Algebraic invariants of a system of matrices , author=. Siberian Mathematical Journal , volume=

  29. [29]

    Advances in Mathematics , volume=

    The invariant theory of n n matrices , author=. Advances in Mathematics , volume=

  30. [30]

    npj Computational Materials , volume=

    High performance Wannier interpolation of Berry curvature and related quantities with WannierBerri code , author=. npj Computational Materials , volume=

  31. [31]

    Existence of a type of optimal norm-conserving pseudopotentials for

    Canc\`. Existence of a type of optimal norm-conserving pseudopotentials for. Communications in Mathematical Sciences , volume=

  32. [32]

    2020 , publisher=

    Electronic structure: basic theory and practical methods , author=. 2020 , publisher=

    2020

  33. [33]

    Magnetic

    Korotyaev, Evgeny and Saburova, Natalia , journal=. Magnetic

  34. [34]

    Artificial intelligence and statistics , pages=

    A linearly-convergent stochastic L-BFGS algorithm , author=. Artificial intelligence and statistics , pages=. 2016 , organization=

    2016

  35. [35]

    2000 , publisher=

    Jacobi operators and completely integrable nonlinear lattices , author=. 2000 , publisher=

    2000

  36. [36]

    2025 , eprint =

    Barrett, Rhyan and Ortner, Christoph and Westermayr, Julia , title =. 2025 , eprint =

    2025

  37. [37]

    Nature Communications , volume=

    Excited state non-adiabatic dynamics of large photoswitchable molecules using a chemically transferable machine learning potential , author=. Nature Communications , volume=

  38. [38]

    2002 , publisher=

    Ordinary Differential Equations , author=. 2002 , publisher=

    2002

  39. [39]

    Reviews of Modern Physics , volume=

    Diabolical conical intersections , author=. Reviews of Modern Physics , volume=

  40. [40]

    Annual Review of Physical Chemistry , volume=

    Understanding the surface hopping view of electronic transitions and decoherence , author=. Annual Review of Physical Chemistry , volume=

  41. [41]

    , title =

    von Neumann, John and Wigner, Eugene P. , title =. Physikalische Zeitschrift , volume =

  42. [42]

    Linear Algebra and Its Applications , volume=

    Multiple eigenvalues , author=. Linear Algebra and Its Applications , volume=

  43. [43]

    2013 , publisher =

    Introduction to Smooth Manifolds , author =. 2013 , publisher =

    2013

  44. [44]

    1973 , publisher =

    Stable Mappings and Their Singularities , author =. 1973 , publisher =

    1973

  45. [45]

    Proceedings of the Royal Society of London

    Berry, Michael Victor , title =. Proceedings of the Royal Society of London. Series A, Mathematical and Physical Sciences , volume =

  46. [46]

    , title =

    Wilczek, Frank and Zee, A. , title =. Physical Review Letters , volume =

  47. [47]

    Thouless, D. J. and Kohmoto, M. and Nightingale, M. P. and den Nijs, M. , title =. Physical Review Letters , volume =

  48. [48]

    Niu, Qian and Thouless, D. J. and Wu, Yong-Shi , title =. Physical Review B , volume =

  49. [49]

    Machine learning configuration-dependent friction tensors in

    Sachs, Matthias and Stark, Wojciech G and Maurer, Reinhard J and Ortner, Christoph , journal=. Machine learning configuration-dependent friction tensors in

  50. [50]

    The Journal of Physical Chemistry C , volume=

    Symmetry-adapted high dimensional neural network representation of electronic friction tensor of adsorbates on metals , author=. The Journal of Physical Chemistry C , volume=